摘要
受传统的图像修复网络启发,提出一种基于深度学习的图像双分支修复算法。将图像的修复过程分为两个部分,一个是修复结构化缺陷部分的全局分支,另一个是修复非结构化缺陷部分的局部分支。首先,用残差模块搭建总体网络。然后,在残差模块前加入压缩激励模块Squeeze-and-Excitation(SE)构建局部分支。最后,引入一种注意力机制上下文转换器块Contextual Transformer(CoT)来统计图像中完好的全局信息以进行图像的结构性缺陷修复,构建全局分支。实验结果表明,该方法修复的图片有0.02的SSIM增益,0.23dB的PSNR提升,优于其他现有的图像修复方法的视觉质量和数值指标,能有效地改善修复性能。
Inspired by traditional image repair networks,a double branch image repair algorithm based on deep learning is pro⁃posed.The process of image repair is divided into two parts,one is the global branch of structural defect repair and the other is the local branch of unstructured defect repair.First,the residual module is used to build the overall network.Then,the compression excitation module Squeeze-and-Excitation(SE)is added to the residual module to construct the local branch.Finally,an attention mechanism Contextual Transformer(COT)is introduced to collect the intact global information in the image for structural defect repair and global branch construction.The experimental results show that the image repaired by this method has 0.02 SSIM gain and 0.23dB PSNR im⁃provement,which is superior to the visual quality and numerical indexes of other existing image restoration methods,and can effective⁃ly improve the restoration performance.
作者
朱立忠
佟昕
ZHU Lizhong;TONG Xin(Shenyang Ligong University,Shenyang 110159,China)
出处
《通信与信息技术》
2024年第5期14-18,55,共6页
Communication & Information Technology
关键词
图像修复
深度学习
双分支修复
注意力机制
Image inpainting
Deep learning
Double branch repair
Attention mechanism